Date of Graduation

5-2020

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science (PhD)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

Wu, Xintao

Committee Member

Li, Qinghua

Second Committee Member

Panda, Brajendra N.

Third Committee Member

Yang, Song

Fourth Committee Member

Zhang, Lu

Keywords

Algorithmic Bias; Causal Inference; Fairness; Machine Learning

Abstract

Fairness is a social norm and a legal requirement in today's society. Many laws and regulations (e.g., the Equal Credit Opportunity Act of 1974) have been established to prohibit discrimination and enforce fairness on several grounds, such as gender, age, sexual orientation, race, and religion, referred to as sensitive attributes. Nowadays machine learning algorithms are extensively applied to make important decisions in many real-world applications, e.g., employment, admission, and loans. Traditional machine learning algorithms aim to maximize predictive performance, e.g., accuracy. Consequently, certain groups may get unfairly treated when those algorithms are applied for decision-making. Therefore, it is an imperative task to develop fairness-aware machine learning algorithms such that the decisions made by them are not only accurate but also subject to fairness requirements. In the literature, machine learning researchers have proposed association-based fairness notions, e.g., statistical parity, disparate impact, equality of opportunity, etc., and developed respective discrimination mitigation approaches. However, these works did not consider that fairness should be treated as a causal relationship. Although it is well known that association does not imply causation, the gap between association and causation is not paid sufficient attention by the fairness researchers and stakeholders.

The goal of this dissertation is to study fairness in machine learning, define appropriate fairness notions, and develop novel discrimination mitigation approaches from a causal perspective. Based on Pearl's structural causal model, we propose to formulate discrimination as causal effects of the sensitive attribute on the decision. We consider different types of causal effects to cope with different situations, including the path-specific effect for direct/indirect discrimination, the counterfactual effect for group/individual discrimination, and the path-specific counterfactual effect for general cases. In the attempt to measure discrimination, the unidentifiable situations pose an inevitable barrier to the accurate causal inference. To address this challenge, we propose novel bounding methods to accurately estimate the strength of unidentifiable fairness notions, including path-specific fairness, counterfactual fairness, and path-specific counterfactual fairness. Based on the estimation of fairness, we develop novel and efficient algorithms for learning fair classification models. Besides classification, we also investigate the discrimination issues in other machine learning scenarios, such as ranked data analysis.

Download

Share

COinS